diff --git a/src/PythonWrapper.cpp b/src/PythonWrapper.cpp
index 2c5e8c2..b45fc46 100644
--- a/src/PythonWrapper.cpp
+++ b/src/PythonWrapper.cpp
@@ -21,6 +21,7 @@
 #include "CProducer.h"
 #include "CPushConsumer.h"
 #include "PythonWrapper.h"
+#include "CMQException.h"
 #include <boost/python.hpp>
 #include <map>
 
@@ -31,6 +32,9 @@
         "PYTHON_CLIENT_VERSION: " PYTHON_CLIENT_VERSION ", BUILD DATE: " PYCLI_BUILD_DATE " ";
 
 map<CPushConsumer *, pair<PyObject *, object>> g_CallBackMap;
+map<CPushConsumer *, pair<PyObject *, object>> g_CallBackMap;
+map<void *, PyObject *> g_SendSuccessCallbackMap;
+map<void *, PyObject *> g_SendExceptionCallbackMap;
 
 class PyThreadStateLock {
 public:
@@ -153,6 +157,29 @@
     return SendMessageOneway((CProducer *) producer, (CMessage *) msg);
 }
 
+void PySendSuccessCallback(CSendResult result){
+    map<void *, PyObject *>::iterator iter = g_SendSuccessCallbackMap.begin();
+    while(iter != pCallbackMap.end()) {
+        boost::python::call<void>(iter->second, iter->first);
+    }
+}
+
+void PySendExceptionCallback(CMQException e){
+    map<void *, PyObject *>::iterator iter = g_SendExceptionCallbackMap.begin();
+    while(iter != pCallbackMap.end()) {
+        boost::python::call<void>(iter->second, iter->first);
+    }
+}
+
+int PySendMessageAsync(void *producer, void *msg, PyObject *sendSuccessCallback , PyObject *sendExceptionCallback){
+    g_SendSuccessCallbackMap[msg] = sendSuccessCallback;
+    g_SendExceptionCallbackMap[msg] = sendExceptionCallback;
+
+    return SendMessageAsync((CProducer *) producer,  (CMessage *) msg, &PySendSuccessCallback, &PySendExceptionCallback);
+}
+
+
+
 PySendResult PySendMessageOrderly(void *producer, void *msg, int autoRetryTimes, void *args, PyObject *queueSelector) {
     PySendResult ret;
     CSendResult result;
@@ -323,6 +350,11 @@
     def("SetProducerInstanceName", PySetProducerInstanceName);
     def("SetProducerSessionCredentials", PySetProducerSessionCredentials);
     def("SendMessageSync", PySendMessageSync);
+
+    def("SendSuccessCallback", PySendSuccessCallback);
+    def("SendExceptionCallback", PySendExceptionCallback);
+    def("SendMessageAsync", PySendMessageAsync);
+
     def("SendMessageOneway", PySendMessageOneway);
     def("SendMessageOrderly", PySendMessageOrderly);
     def("SendMessageOrderlyByShardingKey", PySendMessageOrderlyByShardingKey);
diff --git a/src/PythonWrapper.h b/src/PythonWrapper.h
index 28bfc1e..5fd46af 100644
--- a/src/PythonWrapper.h
+++ b/src/PythonWrapper.h
@@ -22,6 +22,7 @@
 #include "CProducer.h"
 #include "CPushConsumer.h"
 #include "CPullConsumer.h"
+#include "CMQException.h"
 #include <boost/python.hpp>
 
 using namespace boost::python;
@@ -82,7 +83,12 @@
 int PySetProducerSessionCredentials(void *producer, const char *accessKey, const char *secretKey, const char *channel);
 PySendResult PySendMessageSync(void *producer, void *msg);
 int PySendMessageOneway(void *producer, void *msg);
-// PySendResult PySendMessageOrderly(void *producer, void *msg , int autoRetryTimes, PyObject *args, PyObject *callback);
+
+void PySendSuccessCallback(CSendResult result);
+void PySendExceptionCallback(CMQException ex);
+int PySendMessageAsync(void *producer, void *msg, PyObject *sendSuccessCallback , PyObject *sendExceptionCallback);
+
+
 PySendResult PySendMessageOrderly(void *producer, void *msg, int autoRetryTimes, void *args, PyObject *queueSelector);
 PySendResult PySendMessageOrderlyByShardingKey(void *producer, void *msg, const char *shardingKey);
 
diff --git a/test/TestSendMessages.py b/test/TestSendMessages.py
index 7258d70..dc80784 100644
--- a/test/TestSendMessages.py
+++ b/test/TestSendMessages.py
@@ -214,6 +214,21 @@
         DestroyMessage(msg)
         print 'msg id =' + result.GetMsgId()
 
+def send_message_orderly_with_shardingkey(count):
+    key = 'rmq-key'
+    print 'start sending sharding key order-ly message'
+    tag = 'test'
+    for n in range(count):
+        body = 'hi rmq sharding orderly-message, now is' + str(n)
+        msg = CreateMessage(topic_orderly)
+        SetMessageBody(msg, body)
+        SetMessageKeys(msg, key)
+        SetMessageTags(msg, tag)
+
+        result = SendMessageOrderlyByShardingKey(producer, msg, 'orderId')
+        DestroyMessage(msg)
+        print 'msg id =' + result.GetMsgId()
+
 def calc_which_queue_to_send(size, msg, arg): ## it is index start with 0....
     return 0